Causal mediation analysis distinguishes pathways through which a treatment affects an outcome. In linear models, ordinary least squares may leave information in the error distribution unused. We develop a semiparametric framework for joint inference in linear mediation models with treatment-mediator interaction and unspecified error distributions. A stacked estimating system integrates regression-error scores and propagates uncertainty from both regressions and the baseline-covariate distribution through natural-effect decompositions. Under the stated model and regularity conditions, we characterize the natural-effect influence functions and establish an OLS-relative asymptotic variance and local-power ordering. We further develop density-score estimation and joint inference for residual-correlation sensitivity analysis of unmeasured mediator-outcome confounding. Reduced-form regressions and joint uncertainty propagation for their coefficients, residual variances and cross-covariances yield pointwise sensitivity curves and variance-explained contours under explicit working assumptions. Simulations show reduced root mean squared error across several non-Gaussian settings and markedly higher power for a weak indirect effect under asymmetric-mixture errors. The power advantage over ordinary least squares and interaction-adapted Huber comparators persists after independent null calibration. Evaluations also examine interval coverage, numerical success, full sensitivity curves and variance misspecification. An analysis of the randomized JOBS II behavioral intervention illustrates how natural-effect estimates, uncertainty and confounding sensitivity can be interpreted together. Complete estimating equations, a deterministic multi-start algorithm and reproducible numerical output accompany the analysis.
翻译:暂无翻译